# What Is an Agentic Cyberattack? What the Bell Canada Breach Tells Us

> Bell Canada says an 'agentic' attacker stole business data. What agentic cyberattacks are, how they work, and how to defend your org and your own AI agents.

- URL: https://incube.club/blog/what-is-an-agentic-cyberattack-bell-canada
- Published: 2026-11-22
- Updated: 2026-11-22
- Author: Incube Team (Incube, https://incube.club)
- Category: News
- Tags: Cybersecurity, Agentic AI, AI agents, Bell Canada, AI security, Tech news

**Quick answer:** An agentic cyberattack is one where AI agents run several stages of an intrusion themselves, such as reconnaissance, login attempts and data theft, while humans only set goals. Bell Canada said an 'agentic threat actor' stole enterprise customer data on November 17, 2026. The best defences are still basics: phishing-resistant MFA, least privilege, fast patching, monitoring and pre-agreed shutdown decisions.

## Key takeaways

- **An agentic cyberattack** uses AI agents to chain the stages of an intrusion together, with humans only setting goals and checking in.
- **Bell Canada said an 'agentic threat actor'** took enterprise customer contact data on November 17, 2026, but has not explained what made it agentic.
- **Speed is the real shift:** Microsoft saw some Storm-0940 intrusions go from first targeting to privileged cloud access in under three hours.
- **The basics still win:** phishing-resistant MFA, least privilege, fast patching, monitoring and pre-agreed shutdown rights break the attack chain.
- **Your own AI agents are software identities too,** so give each one narrow permissions, human approvals for risky actions and a tested off switch.

On Tuesday, November 17, someone broke into Bell Canada and walked off with business customer data. That part, sadly, is ordinary. The unusual part came three days later, when Bell described the intruder with a word normally found in conference keynotes: **agentic**.

"Bell is investigating a cybersecurity incident that occurred on November 17 in which an agentic threat actor accessed some of our systems," a spokesperson told [The Globe and Mail](https://www.theglobeandmail.com/business/article-bell-cybersecurity-incident-customer-data/). It is an unusual word to find in a breach statement. It probably won't be the last time we see it.

So what is an agentic cyberattack, how worried should you be, and what does it change for the AI agents your own team is starting to run? Let's take it one question at a time.

![Bell Canada's campus building in Montreal with the blue Bell logo on its facade](https://incube.club/blog/images/bell-canadas-campus-building-in-montreal-with-the-57bf3ea74b.webp "Bell Campus in Montreal. Photo: oknidius via Wikimedia Commons, CC BY 2.0")

## What happened in the Bell Canada breach?

Bell says an attacker got into some of its systems on November 17, including one used for ordering and provisioning business circuits, and took enterprise customer data. Here is what the company and the reporting so far say, and nothing more:

- **What was taken:** business contact details such as names, titles, corporate email addresses and phone numbers, plus some IP addresses, according to [CBC News](https://ca.news.yahoo.com/bell-says-investigating-cyberattack-led-215402193.html).
- **What was not:** Bell says no highly sensitive personal information was taken, and files and PDFs were not affected.
- **Where they got in:** [BleepingComputer](https://www.bleepingcomputer.com/news/security/bell-canada-says-agentic-hacker-stole-enterprise-customer-data/) and SC Media report that the intruder reached a pre-production environment of Bell's Maximo system.
- **Impact on service:** none, says Bell. It contained the incident and secured its systems.
- **Who did it:** unknown. Nobody has claimed it.
- **Next steps:** Bell is notifying affected customers and sharing indicators of compromise with them and with the Canadian Centre for Cyber Security. It has also contacted the RCMP.

Notice what is missing. Bell has not explained what made the attacker "agentic." No technical evidence has been published yet. Keep that in mind; we'll come back to it.

## What is an agentic cyberattack?

An agentic cyberattack is an attack in which AI agents carry out several stages of the intrusion themselves, with humans setting the goal and checking in only now and then. Reconnaissance, login attempts, moving around the network and pulling data out get chained together and run by software.

The [Canadian Centre for Cyber Security put it neatly](https://www.cyber.gc.ca/en/guidance/cyber-threat-bulletin-agentic-artificial-intelligence-accelerating-cyber-attacks) in an October bulletin: AI already made it easier to find weaknesses and write malicious code, but "agentic AI goes further by linking separate stages of an attack and carrying them out with less human direction."

That is the key idea. None of the individual techniques are new. Password spraying is old news. What is new is a tireless intern that does every step, in order, across hundreds of targets, without needing coffee or a pep talk.

## Was the Bell attacker really an AI agent?

We don't know yet, and anyone who tells you otherwise is guessing. Bell used the word but did not explain it. Reporters at the Globe noted the term "suggests that artificial intelligence was involved," which is fair, and also all we have.

There are two honest readings:

1. **An attacker used an agent framework.** This is the pattern seen in recent, well-documented campaigns (more below). A human picks the target; agents do the legwork.
2. **"Agentic" describes speed and automation more loosely.** Companies sometimes reach for the newest word to describe an attack that moved unusually fast.

Either way, the lesson for defenders is the same. Plan as if the next attacker is a machine that works at machine speed.

## Why are agentic attacks faster and harder to stop?

Agentic attacks are faster because the slow, human parts in the middle of an intrusion are gone. Two public cases show how much that matters.

**Case 1: GTG-1002 (2025).** In November 2025, Anthropic published [a report on a campaign it detected in mid-September](https://www.anthropic.com/news/disrupting-AI-espionage). It says a group it assessed as Chinese state-sponsored used Claude Code to attempt intrusions into roughly thirty organisations, succeeding in a small number. Anthropic estimated the AI did 80 to 90 percent of the work. Humans stepped in at only a handful of decision points. The attackers got past safeguards by breaking the job into small, innocent-looking tasks and claiming to be a security firm doing defensive testing. One comforting detail: the AI sometimes hallucinated credentials, or "found" secrets that were already public.

**Case 2: Storm-0940 and ARTEX (2026).** On October 21, [Microsoft Threat Intelligence described](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/) a China-state aligned actor using ARTEX, an open-source penetration testing framework that can plug into any AI model. Microsoft says it was used against hundreds of organisations. In several cases, the time from first targeting to privileged access in the cloud was **under three hours**. After the actor took control of identities belonging to ConnectWise cloud applications, its agents tried to sign in to more than 2,250 customer tenants using the trust those apps already had. Microsoft also notes that most of the targeting failed.

Here is what changes when the attacker is an agent:

| | Human-led intrusion | Agentic intrusion |
|---|---|---|
| Who does the middle steps | People, one target at a time | Agents, many targets in parallel |
| Speed from foothold to admin | Often days or weeks | Under three hours in some Storm-0940 cases |
| Scale | Limited by staff | Limited mostly by compute |
| Human role | Hands on keyboard | Sets goals, approves key steps |
| Typical weak spot used | Weak passwords, over-broad access, unpatched systems | The same ones, found faster |

That last row is the good news hiding in the bad news.

![Timeline of agentic cyberattack milestones from Anthropic's September 2025 detection to the Bell Canada incident on November 17, 2026](https://incube.club/blog/images/timeline-of-agentic-cyberattack-milestones-from-anthropics-s-2f757036e5.webp "From the first documented AI-run campaign to Bell Canada in about 14 months. Sources linked in the article.")

## How do you defend against an agentic cyberattack?

You defend against an agentic attack mostly with the same controls that stop human attackers, applied more strictly and with faster decisions. The Cyber Centre is blunt about it: "Agentic AI does not make established cyber security practices obsolete. It makes them more important."

Its October bulletin recommends that organisations:

1. **Require phishing-resistant MFA** for administrators, remote access and other important accounts.
2. **Limit access** for users, applications and service accounts.
3. **Review third-party and shared cloud apps** and what they can reach.
4. **Patch internet-facing systems fast** and switch off services nobody uses.
5. **Keep logs and watch them** for odd sign-ins, new credentials and sudden changes in access.
6. **Decide in advance who can pull the plug** on an account, app, network or service.
7. **Practise incidents that unfold in hours,** not days.
8. **Separate important systems** and keep tested offline backups.

Point 6 deserves a sticky note on every CISO's monitor. The bulletin warns that "decisions that once allowed time for discussion may soon have to be made within minutes." If your incident plan begins with "schedule a call," an agent will have finished before the calendar invite lands.

![Diagram of an agentic attack chain: find targets, test logins, explore the network, grab more access, take the data, with a matching defence under each step](https://incube.club/blog/images/diagram-of-an-agentic-attack-chain-find-targets-d24b2a7c37.webp "Every stage of the chain has a boring control that breaks it. Defences from the Canadian Centre for Cyber Security bulletin.")

The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/), published in December 2025 with input from more than 100 experts, is a good companion read. It lists risks such as agent behaviour hijacking, tool misuse and identity and privilege abuse.

## What does this mean for the AI agents your own team runs?

Here is the part most coverage skips. The Storm-0940 story is not only about attackers using agents. It is about **trusted software identities being turned against the companies that trusted them.** Those ConnectWise apps already had permission inside thousands of tenants. Nobody had to log in as a person.

Your AI agents are software identities too. An agent that reads your shared inbox, posts to your ERP or replies to customers holds credentials and standing trust. That makes it useful. It also makes it a target.

The Five Eyes agencies' joint guidance, [Careful adoption of agentic AI](https://www.cyber.gc.ca/en/guidance/careful-adoption-agentic-ai), says it plainly: never give agentic AI "broad or unrestricted access, especially to sensitive data or critical systems."

When we build agents at Incube, this is where the design starts, before any prompt is written. Take an [invoice processing agent](https://incube.club/work/invoice-processing-agent). We ask four questions:

1. **What is the smallest set of actions it needs?** It can read invoices and draft entries. It cannot pay anyone.
2. **Which credentials does it hold, and for how long?** Scoped, short-lived and rotated, never a shared admin key.
3. **What does a human approve?** Anything that moves money or changes a supplier's bank details goes to a person.
4. **How fast can we stop it?** Every agent has a tested off switch and an owner who is allowed to use it. (We wrote a whole guide on [AI agent kill switches](https://incube.club/blog/what-is-an-ai-agent-kill-switch).)

This is why we keep arguing for specialised agents over one do-everything assistant. A narrow agent with narrow permissions is easier to test, easier to verify and easier to monitor. If its credentials leak, the blast radius is small. A general agent with the keys to everything is the same risk as that over-privileged multi-tenant app, just with better manners.

The flip side is also true. Defenders need machine speed now. Bell's own security arm has been building AI tools for investigations, and the Cyber Centre's warning about decisions "within minutes" is really an argument for well-governed agents on defence: agents that triage alerts, pull logs and draft the containment step, while a human makes the call. The [inbox triage agent](https://incube.club/work/inbox-triage-agent) pattern, sort, check, escalate, works for security queues as much as for customer email.

## Is this the start of autonomous hacking?

Not quite, and it is worth being precise. The Cyber Centre says attacks "are not yet fully autonomous": humans still pick targets, set objectives and oversee the work. Anthropic's report shows the AI still makes mistakes. Microsoft says most Storm-0940 targeting failed.

What has changed is the middle of the attack. That middle used to take a team days. Now it can take software a few hours. The Cyber Centre also expects AI-enabled attacks to become "more targeted, less visible and harder to detect" as the tools improve. So the window to fix the basics is open, but it is not open forever.

## What should you ask your team this week?

Start with one question: **"If an agent got hold of one of our service accounts or AI tool credentials tonight, what could it reach by breakfast?"**

If the honest answer is "we'd have to check," you have your first project. List every app, integration and AI agent that holds credentials. Cut each one down to what it truly needs. Name the person who can switch each one off, and make sure they know it.

If you are planning your first agents, our guide on [how to choose your first AI agent](https://incube.club/blog/how-to-choose-your-first-ai-agent) helps you pick work that is high value and low risk. And if you'd rather have a specialised agent built, tested and monitored for you with humans firmly in control, that is what we do at Incube. [Start a conversation here](https://incube.club/#contact).

## Frequently asked questions

### What is an agentic cyberattack?

An agentic cyberattack is an intrusion where AI agents carry out several stages on their own, such as scanning for targets, testing passwords, exploring a network and taking data. Humans still choose targets and objectives, but the software links the steps together and runs them quickly across many organisations. The techniques are familiar; the speed and scale are new.

### What happened in the Bell Canada breach in November 2026?

Bell Canada says an 'agentic threat actor' accessed some of its systems on November 17, 2026, including one used to order and provision business circuits, and took enterprise customer data such as names, titles, work emails, phone numbers and some IP addresses. Bell says services were not affected, sensitive files were not taken and the attacker has not been identified.

### How fast can an AI-driven cyberattack move?

Very fast. Microsoft reported in October 2026 that the Storm-0940 group, using the ARTEX agent framework, went from first targeting to privileged cloud access in under three hours in multiple cases. Canada's Cyber Centre warns that security decisions that once allowed time for discussion may soon need to be made within minutes.

### How can a company defend against agentic AI attacks?

Mostly with proven basics applied strictly: phishing-resistant multi-factor authentication, least-privilege access for people, apps and service accounts, fast patching of internet-facing systems, good logging and monitoring, separated critical systems with tested offline backups, and a pre-agreed decision on who can isolate an account or service. Practise incidents that unfold in hours, not days.

### Are AI agents used inside a business a security risk?

They can be, because every agent holds credentials and standing trust. The safe pattern is specialised agents with the smallest set of permissions they need, short-lived credentials, human approval for risky actions such as payments, full monitoring and a tested off switch. Incube builds agents this way so a leaked credential has a small blast radius.
